lenet1 model精度 聚类个数 聚类算法 50 60 70 80 90 100 110 120 130 140 150 160 170 180
    lenet1 PACE(paper) 0.9486 0.742 0.222 0.494 1.110 0.295 0.190 0.315 0.181 0.163 0.574 0.158 0.202 0.434 0.385
    PACE(recur1) 0.982 0.140 0.494 0.140 0.416 0.801 1.282 0.742 0.967 1.289 0.900 0.520 0.708 0.971
    PACE(recur2) 0.982 0.222 0.574 0.140 0.416 0.801 1.224 0.742 1.061 0.534 0.158 0.109 0.434 0.696
    SRS 3.455 2.928 2.380 2.088 2.494 2.674 1.851 2.035 1.794 1.630 1.905 1.732 1.792 1.711
    m2(805) 15 MiniBatchKMeans(bactch_size=200) 2.860 1.640 0.574 0.321 0.696 0.089 0.595 0.098 0.204 0.574 0.193 0.077 0.154 0.112
    10 KMEANS 3.023 3.193 3.431 2.360 1.602 0.921 0.365 0.181 0.286 1.335 0.860 0.485 0.154 0.385
    15 KMEANS 3.193 3.057 1.902 2.360 1.453 0.921 1.224 0.742 0.163 0.175 0.442 0.737 0.708 0.324
    30 KMEDOID 0.742 0.055 0.854 1.343 1.731 1.058 1.402 0.693 0.329 0.104 0.504 0.765 0.378 0.416
    40 KMEDOID 1.243 0.055 0.792 1.390 1.731 0.929 1.436 0.012 0.378 0.007 0.345 0.710 0.350 0.620
    50 KMEDOID 1.110 0.123 0.728 1.244 1.692 0.885 0.467 0.742 0.416 0.068 0.339 0.044 0.884 0.574
    lenet4 PACE(paper) 0.9679 1.249 1.790 1.076 0.494 0.087 0.240 0.483 0.731 0.920 1.082 0.561 0.710 0.871 0.988
    PACE(recur) 1.210 1.798 1.138 0.540 0.123 0.210 0.458 0.710 0.902 1.067 1.210 0.694 0.857 0.975
    SRS 2.423 1.881 1.981 1.787 1.770 1.606 1.582 1.859 1.449 1.527 1.313 1.678 1.252 1.208
    m2(1181) 15 MiniBatchKMeans(bactch_size=200) 0.712 0.180 0.393 0.710 1.036 1.210 0.483 0.731 0.920 1.036 0.123 0.564 0.340 0.142
    30 KMEANS 0.872 3.570 3.832 2.888 2.225 1.841 1.335 0.888 0.578 0.387 0.101 0.564 0.383 0.123
    30 KMEDOID 3.040 1.875 2.504 1.853 1.284 0.750 0.426 0.069 0.133 0.290 0.543 0.741 0.871 1.000
    m3(1181)_scale 20 KMEANS 1.169 1.708 2.424 1.790 1.284 1.740 1.335 0.888 0.607 0.361 0.168 0.045 0.299 0.105
    m3(1181)_scale 30 KMEANS 3.210 1.486 1.761 0.771 1.234 0.790 0.426 0.069 0.085 0.361 0.123 0.025 0.319 0.657

    svhn PACE(paper) 0.8790 1.578 0.164 1.578 1.639 0.646 0.676 0.070 0.008 0.164 0.396 0.968 0.779 0.612 1.082
    PACE(recur) 1.390 0.017 1.229 0.340 1.578 0.396 1.931 0.904 0.678 1.182 1.712 2.093 1.849 2.260
    SRS 5.477 3.323 3.582 3.154 4.061 2.956 2.955 2.591 2.626 2.970 2.512 2.556 2.442 2.361
    m2(629) 16 MiniBatchKMeans(bactch_size=200) 4.222 2.650 2.602 2.896 1.379 0.641 0.045 0.292 1.074 1.276 0.734 0.376 0.795 0.016
    Fashion PACE(paper) 0.8988 0.316 0.284 1.669 1.478 2.428 2.199 1.111 0.203 1.330 0.518 1.801 1.060 0.406 0.175
    PACE(recur)
    SRS 3.691 4.847 3.823 3.667 2.974 3.209 3.142 2.388 2.477 2.590 2.688 2.175 1.980 2.181
    m2(1809) 15 MiniBatchKMeans(bactch_size=200) 8.159 6.787 5.772 2.620 3.453 2.199 2.847 1.717 0.818 0.700 0.724 0.803 0.178 0.676
    resnet20 PACE(paper) 0.9145 2.668 0.353 0.099 1.143 0.241 1.351 2.261 1.367 1.374 1.379 1.384 0.767 1.976 1.947
    PACE(recur1) 0.9145
    PACE(recur2) 0.9122
    SRS 0.9122 3.897 3.279 3.118 3.012 3.170 2.578 2.242 2.637 2.236 2.355 2.148 1.858 2.527 2.233
    m2 0.9122 15 0.780 2.000 1.534 1.280 0.109 0.220 0.598 0.311 0.253 0.084 0.672 0.705 0.545 0.400
    lenet5 PACE(recur2) 0.9872 1.280 1.280 1.280 1.280 1.280 1.280 1.280 1.280 0.517 0.566 0.613 0.659 0.688 0.721
    m2 0.9872 15 MiniBatchKMeans(bactch_size=200) 0.681 0.359 0.169 0.030 0.156 0.280 0.387 0.454 0.517
    聚类方法 k的范围 确定k的方法 batch_size bestK 50 60 70 80 90 100 110 120 130 140 150 160 170 180
    PACE 2.668 0.353 0.099 1.143 0.241 1.351 2.261 1.367 1.374 1.379 1.384 0.767 1.976 1.947
    resnet20 m2_1357 KMeans 5～50 CH 100 5 1.220 1.220 1.220 0.030 0.209 0.780 1.507 0.377 0.318 0.923 1.447 0.655 0.496 0.447
    5
    		6.780 5.501 5.923 6.280 6.582 5.780 5.144 5.419 4.165 3.780 3.447 3.155 2.309 2.669
    5～50 DB 100 11 2.898 3.695 4.432 2.607 3.162 3.780 4.276 4.613 4.165 4.525 4.806 5.030 4.662 4.847
    5～50 SH 100 5 6.780 5.501 5.881 5.030 4.384 4.820 4.235 4.578 4.165 3.780 3.447 3.120 2.309 2.669
    15～40 CH 100 15 4.945 5.974 4.263 3.720 4.407 4.220 3.947 2.985 3.623 3.986 4.465 3.720 3.573 3.998
    15～40 DB 100 36 2.397 0.149 0.175 1.347 0.311 0.447 1.220 2.985 2.939 3.450 2.404 2.876 2.203 2.209
    15～40 SH 100 23 5.220 4.779 2.488 2.470 1.220 2.220 3.038 2.790 1.989 1.292 2.478 3.021 2.985 2.695
    MiniBatchKMeans 5～50 CH 100 5 4.698 3.695 1.637 1.280 1.002 1.780 1.507 0.516 1.088 1.637 1.447 1.233 0.545 1.002
    5～50 DB 100 15 4.613 2.000 0.209 2.470 2.331 3.101 3.038 2.053 1.220 0.506 1.287 1.283 1.808 1.332
    5～50 SH 100 5 2.780 3.695 3.066 3.717 3.224 1.780 1.507 1.280 1.028 1.534 1.447 1.280 1.721 2.076
    15～40 CH 100 15 2.898 2.113 0.209 1.185 2.038 2.719 3.224 2.057 1.146 0.923 1.447 1.862 1.762 1.002
    50 16 6.906 6.474 6.934 6.220 5.827 4.220 2.932 2.985 2.758 1.934 2.031 1.220 1.339 0.664
    150 15 2.780 3.862 3.066 2.607 1.088 0.044 0.598 1.280 1.749 1.586 1.495 1.862 2.271 1.517
    200 15 0.780 2.000 1.534 1.280 0.109 0.220 0.598 0.311 0.253 0.084 0.672 0.705 0.545 0.400
    250 16 3.220 4.335 3.896 3.878 3.308 2.985 2.229 1.304 1.068 0.573 0.055 0.479 0.577 0.209
    300 16 2.870 3.780 3.146 3.717 4.336 2.719 3.325 2.113 1.857 2.305 2.740 3.155 2.271 2.076
    350 15 7.547 4.553 4.263 3.566 2.456 2.331 3.038 1.964 1.220 3.277 2.553 2.541 2.331 1.717
    400 15 2.898 2.113 1.637 1.280 1.088 0.131 0.598 0.447 0.318 0.147 0.004 0.030 0.044 0.109
    450 16 1.024 0.447 0.209 1.088 1.002 1.849 1.573 1.280 0.253 0.269 0.833 0.705 1.088 1.477
    500 15 0.617 1.220 0.044 1.185 2.038 2.719 1.507 1.463 1.910 1.688 2.113 2.569 2.347 2.669
    20～40 100 20 0.780 0.583 1.427 1.280 1.002 0.859 0.672 1.217 0.906 0.440 0.055 0.030 0.577 0.011
    50 21 4.698 3.695 3.066 2.530 2.187 1.780 1.441 0.447 0.451 0.506 0.616 0.654 1.162 1.835
    150 20 1.088 2.328 1.534 2.607 2.038 0.699 0.229 0.305 1.144 1.292 0.553 1.158 0.744 0.664
    200 20 0.617 0.305 1.079 1.097 0.084 0.617 0.149 0.377 1.146 0.866 1.447 1.373 1.679 1.477
    250 20 1.220 1.056 1.220 1.347 0.209 1.321 1.405 1.220 0.522 0.573 1.287 1.283 1.279 1.220
    300 20 3.011 2.000 0.084 1.280 0.011 0.699 0.598 0.447 0.383 0.209 0.780 0.030 0.545 0.870
    350 20 1.220 0.583 0.209 1.220 2.209 1.220 1.312 1.137 0.451 1.783 1.816 1.845 2.267 1.894
    400 20 0.447 1.056 0.084 0.138 0.011 0.699 1.573 1.280 0.318 0.269 0.113 0.025 0.545 1.045
    450 21 1.088 2.000 1.738 2.370 1.002 0.617 0.394 0.542 0.451 1.149 0.616 0.654 0.744 0.209
    500 21 4.945 5.974 4.077 2.470 2.331 4.091 4.064 3.825 2.848 2.731 2.553 1.979 1.220 1.276
    25～40 50 26 1.424 1.220 1.668 2.470 1.447 1.321 0.311 0.387 0.186 0.506 0.004 0.654 0.522 0.109
    100 25 2.780 2.000 2.983 2.530 3.162 3.678 4.276 2.995 2.578 0.923 1.495 1.905 2.385 1.598
    150 26 1.220 1.565 4.263 4.970 4.856 4.485 4.064 2.985 2.848 2.011 1.287 1.347 1.808 1.717
    200 25 4.698 2.000 1.427 1.280 1.088 0.780 1.507 2.000 1.857 2.258 2.780 3.155 2.309 2.076
    250 27 1.424 1.056 0.329 0.138 0.915 0.699 1.507 0.377 0.380 1.934 1.287 0.479 0.522 0.209
    300 25 0.780 0.583 1.836 1.280 0.825 1.849 0.229 1.304 1.989 1.858 1.887 1.845 1.871 2.209
    350 26 0.780 0.305 1.220 1.476 1.220 1.220 2.129 2.602 2.073 1.858 1.287 1.158 0.577 0.159
    400 25 1.024 0.715 2.649 3.720 3.580 3.341 3.720 2.985 2.157 3.277 3.062 3.095 2.463 1.717
    450 25 3.465 1.220 0.084 1.097 1.332 0.220 0.229 0.464 0.311 0.209 0.064 0.025 1.279 0.612
    500 25 3.220 2.501 2.649 3.878 3.308 2.331 3.257 3.515 2.670 2.011 1.287 1.779 1.685 1.389
    15～40 DB 100 38 2.530 3.695 2.810 2.451 2.038 0.617 0.479 0.542 0.669 0.573 0.553 1.283 1.220 0.717
    20～40 DB 100 21 4.780 3.695 4.432 3.717 3.285 3.678 3.325 3.738 2.626 1.586 0.055 0.030 0.447 0.447
    15～40 SH 100 17 2.658 0.305 0.329 0.081 0.109 0.131 0.523 1.280 1.803 0.979 1.447 1.185 1.088 1.598
    20～40 SH 100 23 2.898 2.113 3.066 2.530 3.162 0.859 0.523 1.280 1.749 2.351 2.820 2.491 1.721 2.076

    🌟k偏小即聚类个数较少时，效果很不稳定（对比第2行和第3行）。
    🌟MiniBatchKMeans相较于KMeans速度快很多，效果也稍微好一点
    🌟CH优于其他两个聚类效果的衡量指标，但实际上也没起什么作用，不如默认聚类个数为15
    🌟kmedoids算法在resnet20上横竖聚完都是一类😵‍💫😵‍💫😵‍💫😵‍💫😵‍💫

    聚类算法：MiniBatchKMeans;
    聚类个数：15
    m3

    50 60 70 80 90 100 110 120 130 140 150 160 170 180
    lenet1 50 0.973 0.140 0.854 1.343 1.769 1.140 1.504 1.779 1.264 1.517 1.086 0.681 0.742 0.904
    100 0.742 3.057 2.106 1.270 2.638 4.860 3.869 3.193 2.612 2.054 1.527 1.033 1.255 0.904
    150 2.860 1.640 0.657 1.033 1.453 0.860 0.265 0.140 0.245 0.068 0.442 0.737 0.378 0.168
    200 3.099 0.055 0.792 0.077 0.696 0.860 1.166 1.527 1.783 1.952 1.527 1.778 1.919 1.564
    250 0.742 0.055 0.574 0.077 1.453 1.931 1.224 0.693 1.062 0.615 0.820 0.520 0.777 0.416
    300 1.058 0.032 0.657 1.033 0.416 1.860 1.224 1.527 1.062 1.289 0.900 0.520 0.708 0.447
    350 1.294 0.222 0.792 1.294 0.646 0.089 0.595 1.008 1.294 0.854 1.166 0.737 0.462 0.696
    400 1.218 1.750 0.494 0.202 0.355 0.921 0.315 0.098 0.329 0.175 0.473 0.737 0.406 0.696
    450 2.860 3.335 2.003 0.958 1.602 2.941 3.042 2.423 2.612 2.717 2.193 1.778 1.331 0.971
    500 2.703 3.335 3.556 3.502 3.749 3.860 3.117 2.360 1.783 1.289 0.900 0.556 1.331 0.937
    lenet4 50 3.210 1.543 0.353 0.540 1.234 0.750 0.394 0.922 0.578 0.311 0.168 0.025 0.216 0.142
    100 3.210 3.210 3.210 1.944 0.963 1.190 1.375 0.731 0.920 1.067 1.197 1.347 1.445 1.543
    150 3.210 3.210 1.802 0.741 1.036 1.230 1.408 1.557 1.683 1.781 1.849 1.944 1.414 1.525
    200 1.249 1.543 1.781 0.678 0.123 0.210 0.458 0.710 0.180 0.353 0.525 0.694 0.857 0.975
    250 3.210 3.210 3.210 1.944 2.099 1.230 1.408 1.577 1.695 1.792 1.868 1.960 2.027 0.988
    300 3.210 3.210 3.210 1.944 0.988 1.210 1.408 1.577 1.683 1.781 1.859 1.952 2.027 1.543
    350 3.210 1.543 0.353 0.678 0.123 0.269 0.507 0.731 0.937 1.067 1.197 0.710 0.871 1.000
    400 3.210 1.543 1.781 0.741 0.087 0.210 0.483 0.731 0.884 1.052 1.183 0.694 0.843 0.975
    450 3.210 1.543 1.781 0.741 0.087 0.210 0.483 0.731 0.884 1.052 1.183 0.694 0.843 0.975
    500 3.210 3.210 3.210 1.960 0.988 1.210 1.424 1.557 1.672 1.771 1.877 1.960 2.027 2.105
    svhn 200 0.104 2.104 0.676 0.396 0.118 0.223 0.740 0.292 0.566 0.048 0.846 1.276 1.453 1.992
    fashion 200 4.120 3.453 3.078 2.620 1.029 0.019 0.028 1.355 1.330 1.937 2.632 2.302 2.089 2.517
    lene5 200 1.280 1.280 1.280 0.014 0.156 0.270 0.363 0.440 0.511 0.561 0.053 0.583 0.454 0.359
    vgg16 200 0.825 2.410 1.495 0.244 0.743 1.002 0.924 1.344 0.487 0.533 0.655 0.618 1.445 1.376

    现在效果不是很稳定，甚至感觉参数其实都没什么意义，举个例子，resnet20，聚类算法为MiniBatchKMeans(batch_size为200)，聚类个数为15，第4行的效果也太差了。

    50 60 70 80 90 100 110 120 130 140 150 160 170 180
    0.780 2.000 1.637 1.373 0.915 0.311 0.311 1.137 0.451 0.209 0.672 0.084 0.593 1.002
    2.985 1.565 2.488 3.415 2.584 2.220 2.229 1.876 2.073 2.090 2.478 1.779 1.279 1.165
    4.698 3.862 4.494 2.530 2.038 2.719 2.416 2.057 1.803 0.866 1.348 1.280 0.545 0.493
    6.906 5.974 5.713 5.144 5.827 5.220 4.856 3.825 5.173 4.695 4.465 3.799 3.646 2.952
    0.937 2.000 1.738 2.530 1.002 0.859 0.447 1.056 1.298 0.440 0.616 1.283 1.339 0.717
    0.780 0.447 0.084 1.280 2.209 1.220 3.038 2.985 2.848 2.011 1.887 1.283 0.632 0.664
    1.424 1.220 2.488 1.347 1.332 1.121 0.311 0.464 1.144 0.573 0.616 0.537 0.008 0.447
    3.220 2.695 1.365 2.331 3.442 2.111 1.312 1.137 1.220 1.220 2.553 2.400 2.463 2.331
    2.758 1.220 2.649 2.612 3.442 3.220 1.934 2.985 1.989 1.934 1.220 2.400 1.746 1.276
    2.658 3.862 2.983 3.842 3.285 3.678 4.276 3.738 3.436 3.066 2.780 3.120 2.898 3.224

    上述的结果其实都是单次的结果（表面看上去是10次取均值），在聚类已经完成然后选取离类中心最近的点导致每次选取样本的差异不会很大。但是如果重新聚类效果就会有很大差别。同样的参数为什么两次聚类的效果能差这么多！

    50 60 70 80 90 100 110 120 130 140 150 160 170 180
    3.687 2.395 2.688 3.522 3.343 2.892 2.607 2.144 2.222 2.513 2.279 2.070 1.821 1.853
    3.476 3.059 3.033 3.467 2.941 2.534 3.064 3.096 2.991 2.925 2.963 2.713 2.715 2.471
    3.571 3.453 3.200 3.077 2.530 2.126 1.812 1.924 1.892 1.751 1.620 1.654 1.640 1.804
    2.547 2.143 2.284 1.993 2.143 2.412 1.745 1.828 1.625 1.741 1.682 1.515 1.648 1.605
    2.013 2.321 2.438 1.919 1.819 1.938 2.037 2.322 2.085 1.885 1.847 1.720 1.424 1.222
    3.295 3.343 2.654 2.569 2.947 2.621 2.457 2.301 2.771 2.337 2.102 2.145 2.373 2.209
    3.564 3.364 2.708 2.840 2.448 2.235 2.098 1.971 2.024 2.036 2.129 2.161 2.154 2.026
    2.558 3.670 3.582 2.645 2.342 1.931 2.285 2.123 1.917 1.938 2.149 2.129 2.034 2.053
    3.119 3.390 4.100 3.341 2.962 2.674 2.883 2.566 2.556 2.313 2.026 2.022 2.139 1.929
    3.117 3.295 2.398 2.431 2.002 2.016 2.062 1.853 1.659 1.441 1.323 1.341 1.145 1.175

    ————————————————————————————————————————-
    对去重之后的矩阵mutate_matrix进行PCA降维(n_components=200)之后效果稳定了很多
    Birch❌
    SpectralClustering❌
    �AgglomerativeClustering❌
    �